Problem casting STL complex to fftw_complex

Question

The FFTW manual says that its fftw_complex type is bit compatible to std::complex class in STL. But that doesn\'t work for me:

<

Answer 1

Re-write your code as follows:

#include <complex>
#include <fftw3.h>
int main()
{
   std::complex<double> x(1,0);
   fftw_complex fx;
   memcpy( &fx, &x, sizeof( fftw_complex ) );
}

Every compiler I've used will optimise out the memcpy because it is copying a fixed, ie at compile time, amount of data.

This avoids pointer aliasing issues.

Edit: You can also avoid strict aliasing issues using a union as follows:

#include <complex>
#include <fftw3.h>
int main()
{
   union stdfftw
   {
       std::complex< double > stdc;
       fftw_complex           fftw;
   };
   std::complex<double> x(1,0);
   stdfftw u;
   u.stdc = x;
   fftw_complex fx = u.fftw;
}

Though strictly this C99 rules (Not sure about C++) are broken as reading from a different member of a union to the one written too is undefined. It works on most compilers though. Personally I prefer my original method.

Answer 2

The idea behind bit-compatibility of fftw_complex and C99 and C++ complex types is not that they can be easily created from one another, but that all functions in FFTW that take pointers to fftw_complex can also take pointers to c++ std::complex. Therefore the best approach is probably to use std::complex<> throughout your program and only convert pointers to these values when calling FFTW functions:

std::vector<std::complex<double> > a1, a2;
....
....
fftw_plan_dft(N, reinterpret_cast<fftw_complex*>(&a1[0]),
                 reinterpret_cast<fftw_complex*>(&a2[0]),
                 FFTW_FORWARD, FFTW_ESTIMATE);
....

Answer 3

reinterpret_cast only works for pointers and references. So you'd have to do this:

#include <complex>
#include <fftw3.h>
int main()
{
   std::complex<double> x(1,0);
   fftw_complex fx(*reinterpret_cast<fftw_complex*>(&x));
}

This assumes that fftw_complex has a copy constructor. To avoid problems with strict aliasing, Goz's solution should be preferred.